Arao AI
Back to search results

Qwen3.5-9B · Q6_K

Start a discussion
NVIDIA DGX Spark128GBllama.cpp v0.49L2 Cross-framework verified

This page aggregates 1 real-world runs of Qwen3.5-9B (Q6_K) on NVIDIA DGX Spark with llama.cpp, contributed by 1 independent source platforms; metrics are averages of published measurements.

110.08 tok/s

Decode

Decode speed

496.59 tok/s

Prefill

Prefill speed

0.49 s

TTFT

Time to first token

21 GB

VRAM

VRAM usage

L2 Cross-framework verified

1

Measured runs

1

Independent sources

Other

Source platforms

2 days ago

Last verified

Performance

  1. llama.cpp · Q6_K (current)Decode 110.08 · Prefill 496.59 ·
  2. Strata · Q3_K_M Decode 209.28 · Prefill 1036.76 ·
  3. llama.cpp · IQ2_M Decode 181.94 · Prefill 684.54 ·
  4. vLLM · IQ2_M Decode 210.69 · Prefill 1114.84 ·
  5. vLLM · Q8_0 Decode 101.73 · Prefill 523.44 ·

Core figures

Decode (avg)
110.08 tok/s
Prefill (avg)
496.59 tok/s
TTFT (avg)
0.49 s
VRAM (avg)
21 GB
MTP acceptance rate
—
TTFB
0.88 GB
Power draw
231.6 W

Configuration

Member-level fields are taken from the most recent run

Model
Qwen3.5-9B
Quantization
Q6_K
Framework
llama.cpp
Version
v0.49
Context length
8192 tokens
Batch size
256
GPU layers
33
Flash Attention
On

Hardware

Nominal and measured figures are shown side by side; whether it runs is the reader's call

GPU
NVIDIA DGX Spark
Nominal VRAM
128 GB
Measured VRAM (avg)
21 GB
OS
Ubuntu 22.04
Driver
565.57
CUDA
12.4
Power draw
231.6 W

Sources & evidence

1 measured records in total, each traceable to its original source

  1. L2 Cross-framework verifiedOtherOriginal link Verified on 2026-10-16

    v0.49 · Ubuntu 22.04 · CUDA 12.4 · 8192 ctx

    110.08 tok/s

    Decode

    496.59 tok/s

    Prefill

    0.49 s

    TTFT

    21 GB

    VRAM

    —

    MTP

    231.6 W

    Power draw

    L2 batch seed 2026-10-16: llama.cpp on hw_id=54 running model_id=49 at Q6_K